Core Concepts
Proposing Active Generalized Category Discovery (AGCD) to address challenges in Generalized Category Discovery (GCD).
Abstract
Generalized Category Discovery (GCD) faces challenges in clustering both old and new classes. AGCD introduces an adaptive sampling strategy, Adaptive-Novel, to select valuable samples for labeling. A stable label mapping algorithm ensures consistent training across different active selection stages. AGCD achieves state-of-the-art performance on various datasets.
Stats
Our method improves the new accuracy of GCD by 25.52%/23.49% on CUB/Air with only ∼ 2.5 samples labeled per class.
In ImageNet-100, our method reduces the gap between old and new accuracy from ∼ 28% to 13.72%.
Our method consistently outperforms others across different scenarios and settings.